FELM

🌐 Website • 🤗 Hugging Face Dataset • 📃 Paper

FELM is a meta benchmark to evaluate factuality evaluation for large language models. The benchmark comprises 847 questions that span five distinct domains: world knowledge, science/technology, writing/recommendation, reasoning, and math. We gather prompts corresponding to each domain by various sources including standard datasets like truthfulQA, online platforms like Github repositories, ChatGPT generation or drafted by authors.

We then obtain responses from ChatGPT for these prompts. For each response, we employ fine-grained annotation at the segment level, which includes reference links, identified error types, and the reasons behind these errors as provided by our annotators.

Download

• Method 1: Download the whole dataset by:

  wget https://huggingface.co/datasets/hkust-nlp/felm/blob/main/all.jsonl
  

• Method 2: Load the dataset using Hugging Face datasets:

  from datasets import load_dataset
  dataset=load_dataset(r"hkust-nlp/felm",'wk')
  print(dataset['test'][0])

Data Description

Dataset Snapshot

Category	Data
Number of Instances	847
Number of Fields	5
Labeled Classes	2
Number of Labels	4427

Descriptive Statistics

Statistic	All	world_knowledge	Reasoning	Math	Science/tech	Writting/Recommendation
Segments	4427	532	1025	599	683	1588
Positive segments	3642	385	877	477	582	1321
Negative segments	785	147	148	122	101	267

Data Fields

Field Name	Field Value	Description
index	Integer	the order number of the data point
source	string	the prompt source
prompt	string	the prompt for generating response
response	string	the response of ChatGPT for prompt
segmented_response	list	segments of reponse
labels	list	factuality labels for segmented_response
comment	list	error reasons for segments with factual error
type	list	error types for segments with factual error
ref	list	reference links

Typical Data Point

{"index": "0", 
 "source": "quora", 
 "prompt": "Which country or city has the maximum number of nuclear power plants?", 
 "response": "The United States has the highest number of nuclear power plants in the world, with 94 operating reactors. Other countries with a significant number of nuclear power plants include France, China, Russia, and South Korea.",
 "segmented_response": ["The United States has the highest number of nuclear power plants in the world, with 94 operating reactors.", "Other countries with a significant number of nuclear power plants include France, China, Russia, and South Korea."], 
 "labels": [false, true],
 "comment": ["As of December 2022, there were 92 operable nuclear power reactors in the United States.", ""], 
 "type": ["knowledge_error", null], 
 "ref": ["https://www.eia.gov/tools/faqs/faq.php?id=207&t=3"]}

Evaluation on FELM

Environment requirements:

transformers 4.32.0
openai 0.27.8
tenacity 8.2.2
tokenizer 3.4.2
pandas 2.0.3

To reproduce our results:

cd eval
#put "all.jsonl" here (Downloaded by Method 1)
bash eval.sh
#You can choose "vicuna_30B", "gpt-3.5-turbo" and "gpt-4" for the parameter "model".
#You can choose "raw", "cot", "link", "content" and "cot-cons"(cot-cons means cot self-consistency method) for the parameter "method".
#replace 'Your OPENAI KEY' with your openai api key if using GPT-3.5 or GPT-4

LEADBOARD (in segment level)

Model	F1 score	Balanced accuracy
GPT4	48.3	67.1
Vicuna-33B	32.5	56.5
ChatGPT	25.5	55.9

We only report the highest scores in this table.

Licenses

This work is licensed under a MIT License.

The FELM dataset is licensed under a Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International License.

Citation

Please cite our paper if you use our dataset:

@inproceedings{
chen2023felm,
title={FELM: Benchmarking Factuality Evaluation of Large Language Models},
author={Chen, Shiqi and Zhao, Yiran and Zhang, Jinghan and Chern, I-Chun and Gao, Siyang and Liu, Pengfei and He, Junxian},
booktitle={Thirty-seventh Conference on Neural Information Processing Systems Datasets and Benchmarks Track},
year={2023},
url={http://arxiv.org/abs/2310.00741}
}